Delayed Supervision for Test-Time Language Models
Delayed semantic supervision as a practical outer training objective for usable test-time memory is supported, while leaving open how much of the benefit derives specifically from delay rather than general question-answering and answer-termination supervision.